refactor(chat): extract SkillTagBuffer from chat_completion stream (A-6, PR-3D-c)

[AVADSA25]

Summary

Final piece of the PR-3D split — extracts the streaming <think> + [SKILL:...] tag-machine out of codec_dashboard.chat_completion._stream_gen into a new, tested module, exactly as the audit recommended.

New codec_chat_stream.py

• SkillTagBuffer — the stateful token processor. feed(token) / finish() are generators that yield clean text fragments to emit; it strips <think>…</think> across chunks and buffers [SKILL:name:query] tags char-by-char so a raw tag never leaks, resolving complete tags via an injected resolve_skill_tag(raw) callback (skill execution is I/O, hence injected). visible_chars lets the caller detect the all-tags-dropped blank-bubble case.
• SKILL_TAG_RE — the shared tag pattern (buffer detection + the dashboard resolver).

_stream_gen now keeps only the SSE/HTTP plumbing (POST, iter_lines, data:/[DONE] framing, keepalive, blank-bubble fallback) + the injected _resolve_skill_tag (budget + allowlist + dispatch). chat_completion 466 → 379 LOC.

Behavior preserved exactly

Including the subtle quirks: a same-chunk </think> is dropped (token zeroed after <think>); think-adjacent text is emitted but not counted toward visible_chars; dropped (resolved-to-empty) tags still emit their empty frame; the 5000-char safety cap; cross-chunk tag assembly. The non-streaming post-LLM [SKILL:] path is untouched.

Bonus: SkillTagBuffer is the tested unit the deferred A-12 dashboard-stream migration needs — the dashboard stream can now consume codec_llm.stream()'s raw tokens through it.

Test plan

• tests/test_chat_stream.py — 13 tests: passthrough; <think> cross-chunk + same-chunk-drop quirk; tag resolved / dropped (no leak) / assembled-across-tokens; non-tag bracket passthrough; prefix divergence; 5000-char cap; finish() flush; regex.
• Full suite: 1464 passed, 23 known-baseline failures, zero new, 74 skipped (chat-path suites — step_budget, agents — all green).
• Ruff: new module + test clean; codec_dashboard.py F-delta vs origin/main = 0.
• No skills/ touched → no manifest regen.
• Manual (Mac Studio): a streaming chat reply renders; a [SKILL:...] the LLM emits resolves inline (no raw tag leak); an all-dropped-tags response shows the fallback bubble.

PR-3D complete

All three monoliths decomposed — A-7 Agent.run (#72), A-5 _dispatch_inner (#73), A-6 chat_completion (this PR).

🤖 Generated with Claude Code